Robotics StackExchange | Archived questions

Obstacle avoidance with pure Vision approach?

Hi

I would like to do obstacle avoidance using pure Vision approach. I will use ROS and Ubuntu as a framework. The application is in construction industry so would like to detect the material winch is handling by the crane to perform obstacle avoidance . My question is , if its possible to use only vision approach for obstacle avoidance where the distance form the camera to the object can be 30-40m. I will use the zoom for better approach. If its possible to use only Vision camera any starting algorithm to look at it?

Thanks

Asked by Astronaut on 2018-02-13 22:23:09 UTC

Comments

Well stereo camera setup is pure vision based , yes you can construct a disparity map which could then be used for depth . Depth can be used for the purpose of mapping while performing path planing keeping obstacle avoidance in mind .

Asked by chris_sunny on 2018-02-14 00:04:46 UTC

https://www.stereolabs.com/

Check out this stereo camera , cost wise its a little on the higher side , but a lot can be achieved using this device .

Asked by chris_sunny on 2018-02-14 00:06:46 UTC

I know this ZED camera. But the distance is quite long, as for some high rise buildings can be over 50 m easily and the depth information from the camera in that case can not be used. So is the distance a problem for pure vision approach?

Asked by Astronaut on 2018-02-14 01:14:16 UTC

What do you mean by "pure vision based approach"? No 3D information?

Asked by gvdhoorn on 2018-02-14 02:56:40 UTC

I mean only using Machine Vision Camera, no other sensors.

Asked by Astronaut on 2018-02-14 04:01:24 UTC

You need good contrast in order for stereo cameras to give you good depth information.

Asked by Kurt Leucht on 2018-02-14 14:28:20 UTC

but if the distance is far , let say more then 50m how can be provided the depth information?

Asked by Astronaut on 2018-02-14 21:32:38 UTC

And is there any code using ROS, Neural Networks or Deep Learning, Machine Vision based obstacle avoidance code? Just o start it as Im beginner in this

Asked by Astronaut on 2018-02-14 21:51:50 UTC

There is Octomap that uses Stereo camera input to generate 3D representation of the world. MoveIt! has a good pipeline that bundles a Motion Planner to Octomap representation. But you need to resolve the distance issue.

Asked by tahsinkose on 2018-09-14 00:00:17 UTC

yes I know. Stereo system will not gonna work by distances like 40-45m.

Asked by Astronaut on 2018-09-14 01:41:44 UTC

unless the 2 cameras are separated by a good distance, right? Isn't it just simple geometry?

Asked by Kurt Leucht on 2020-04-03 13:34:36 UTC

Answers